Research map / Adversarial robustness and security / Certified and empirical robustness
Vulnerability-Aware Robust Multimodal Adversarial Training
Junrui Zhang, Xinyu Zhao, Jie Peng, Chenjie Wang, Jianmin Ji, Tianlong Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add VARMAT gradient-norm regularization | FGSM-RS · CMU-MOSEI PGD lambda=0.5 | robust accuracy: 55.76 -> 68.49 | Improved |
| add VARMAT | FGSM-PCO · UR-FUNNY PGD lambda=0.5 | robust accuracy: 41.21 -> 59.55 | Improved |
| add VARMAT | FGSM-EP · AVMNIST PGD lambda=0.5 | robust accuracy: 0.35 -> 11.08 | Improved |
| add VARMAT | FGSM-EP · CMU-MOSEI | clean accuracy: 79.09 -> 78.07 | Hurt |
| add VARMAT | FGSM-RS · UR-FUNNY | clean accuracy: 65.69 -> 63.80 | Hurt |
| add VARMAT | FGSM-RS · AVMNIST | clean accuracy: 69.38 -> 63.21 | Hurt |
| add VARMAT | FGSM-RS · CMU-MOSEI | clean accuracy: 78.83 -> 79.15 | Improved |
| regularize all modalities instead of only the most vulnerable (vs S1) (ablation) | VARMAT FGSM-RS · CMU-MOSEI PGD lambda=0.5 | robust accuracy: 62.63 -> 68.49 | Improved |
| regularize instead of reallocating perturbation budget in attack (vs S2) (ablation) | FGSM-RS · CMU-MOSEI PGD lambda=0.5 | robust accuracy: 35.54 -> 68.49 | Improved |
| regularize all modalities instead of attacking only the most vulnerable (vs S3) (ablation) | FGSM-RS · CMU-MOSEI PGD lambda=0.5 | robust accuracy: 51.88 -> 68.49 | Improved |
| exclude feature norm from regularizer (vs Trap) (ablation) | FGSM-PCO · input-space PGD lambda=0.1 | robust accuracy: 2.76 -> 53.67 | Improved |
| exclude feature norm from regularizer (vs Trap) (ablation) | FGSM-RS · CMU-MOSEI feature PGD lambda=0.5 | robust accuracy: 75.83 -> 68.49 | Hurt |
Closest papers by meaning
- Defending Multimodal Fusion Models Against Single-Source Adversaries (CVPR 2021 · similarity 0.91)
- On the Robustness of Large Multimodal Models Against Image Adversarial Attacks (CVPR 2024 · similarity 0.89)
- Quantifying and Enhancing Multi-modal Robustness with Modality Preference (ICLR 2024 · similarity 0.87)
- Enhance Modality Robustness in Text-Centric Multimodal Alignment with Adversarial Prompting (AAAI 2025 · similarity 0.87)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.86)
- Boosting Adversarial Robustness with CLAT: Criticality Leveraged Adversarial Training (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
